Get email delivery of the Cadence blog featured here
CDNLive India took place a few weeks ago and we are just trying to catch our breath! If you missed it, I'm going to be posting two cool videos before the weekend with the highlights.
Here are two blogs by the veteran blogger Paul McLellan - one on Asynchronous Design and the other a general one on the overall event.
We had several excellent keynotes at CDNLive India and I will be blogging about them in the next few weeks. Today’s blog is following on from my earlier blog in which I had written about a cool new technology that Dr Avneesh Agrawal’s company, Netrydyne, is developing, which is a combination of IoT and Edge Computing. Avneesh was one of the keynote speakers at CDNLive India.
Avneesh started by setting the context - the whole idea is traditional IoT devices is that you have a sensor with a low power processor connected to the cloud, you gather the information, update to the cloud and use some intelligent analytics. But now, he said, we are reaching a point with silicon and low power silicon that we can move a lot of processing to the Edge.
To give some background - when you start moving the processing to the Edge, the amount of data that you can crunch is orders of magnitude higher than transferring the data to the cloud. And that presents a whole suite of applications that can mine this data. One of them is HD mapping, which is what this blog is about.
Avneesh said that the autonomous stack can generally be broadly considered as having two parts: one is the perception, i.e., analyzing the points of interest around you; and the second is route planning. Netradyne is closely associated with the perception part of the stack.
The value of the stack really depends on how many miles you’ve driven. All the Driveri devices that have been deployed are collecting data from driving 8 million miles a month in the US; given the rate of growth Netradyne expects do to about a 100 million miles a month and if things go according to plan, in a few years they will be doing a billion miles a month. The key here is that the Driveri systems will be collecting rich vision-based data, training their systems, capturing the entire semantic representation (all at the Edge), and then uploading it to the cloud for further processing.
At this point you may be saying, so what? Well, the “so what” is that this kind of “perception mapping” is critical to HD maps. A little background: HD maps are used for all autonomous cars, from Level 2 to Level 5, to aid in driving conditions, to understand the road environment. The interesting thing is that these maps need to be updated dynamically, because lanes could change, you could have construction zones, accidents, objects on the road, etc., and so the faster you can update the maps, the lower the risk of an autonomous car getting into…shall we say…a tricky situation.
One of the challenges of HD mapping is that often, especially in India, the road geometry is not obvious to a machine (sometimes not even to a human!). There are no lane lines marked, there could construction zones that pop up arbitrarily, unplanned road closures (due to festivals, for example), to name just a few. What it means is that you may have to do a lot of manual review to create HD maps using traditional systems. And these systems could end up being prohibitively expensive if you add up driver, fuel, and car maintenance costs.
This is where the Netradyne HD maps solution comes in. With a combination of stereoscopic vision and crowdsourcing, it can create very accurate, 3D hi-definition maps. Crowdsourcing is imperative so that you get an accurate picture of the roads. With single sourcing the maps could be inaccurate.
Netradyne already has thousands of vehicles driving around with Driveri, collecting data about the same environment with slightly different angles, so crowdsourcing is happening all the time. It’s a collateral benefit of the Driveri system. “Using SLAM, which is basically localization techniques within the visual scene, we can take all of these different views of the environment, stitch it together and essentially create a very accurate 3D model,” said Avneesh.
Another collateral benefit is that thanks to the use already deployed systems, the cost of generating these maps is orders of magnitude less than the traditional system of sending out fleets of cars and drivers. In fact, Avneesh estimates that Netradyne can map the entire US in less than $10 million using Driveri!
Avneesh concluded by saying that by applying IoT, especially with Edge-based computing, and of course future technologies (including less expensive LIDAR systems), there will be no specialised equipment and there will be no human review like what is required for creating maps today. Highly accurate maps can be developed automatically using computer vision, sensors and crowdsourcing, much like Netradyne is doing now.
In addition, IoT with Edge computing offers a huge opportunity to get access to data there was simply no access to earlier. There are multiple ways to mine this data to create value – HD mapping is just one market.